gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[BACKEND] Support dot op when the output is mma encoding and allowtf32 is true (#937)

Browse Source
This commit is contained in:
Keren Zhou
2022-12-03 11:14:12 -08:00
committed by GitHub
parent 8edfe813a5
commit f2fcaeabf3
5 changed files with 105 additions and 72 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
test/Conversion/tritongpu_to_llvm.mlir
View File

@@ -923,6 +923,45 @@ module attributes {"triton_gpu.num-warps" = 4 : i32} {
// -----
#mma = #triton_gpu.mma<{version=2, warpsPerCTA=[2, 2]}>
#shared = #triton_gpu.shared<{vec = 1, perPhase = 1, maxPhase = 1, order = [1, 0]}>
#blocked = #triton_gpu.blocked<{sizePerThread = [1, 4], threadsPerWarp = [2, 16], warpsPerCTA = [1, 4], order = [1, 0]}>
#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#mma}>
#dot_operand_b = #triton_gpu.dot_op<{opIdx=1, parent=#mma}>
module attributes {"triton_gpu.num-warps" = 4 : i32} {
// CHECK-LABEL: matmul_tf32dot
func @matmul_tf32dot(%ptr:!tt.ptr<f32> {tt.divisibility = 16 : i32},
%a:tensor<32x16xf32, #shared>, %b:tensor<16x32xf32, #shared>) {
%cst = arith.constant dense<0.000000e+00> : tensor<32x32xf32, #mma>
// CHECK: llvm.inline_asm
// CHECK-SAME: ldmatrix.sync.aligned.m8n8.x4.shared.b16
// CHECK-SAME: (f32, f32, f32, f32)
// CHECK: llvm.inline_asm
// CHECK-SAME: ldmatrix.sync.aligned.m8n8.x4.shared.b16
// CHECK-SAME: (f32, f32, f32, f32)
%a_mat = triton_gpu.convert_layout %a : (tensor<32x16xf32, #shared>) -> tensor<32x16xf32, #dot_operand_a>
%b_mat = triton_gpu.convert_layout %b : (tensor<16x32xf32, #shared>) -> tensor<16x32xf32, #dot_operand_b>
// CHECK: llvm.inline_asm
// CHECK-SAME: mma.sync.aligned.m16n8k8.row.col.f32.tf32.tf32.f32
// CHECK: llvm.inline_asm
// CHECK-SAME: mma.sync.aligned.m16n8k8.row.col.f32.tf32.tf32.f32
// CHECK: llvm.inline_asm
// CHECK-SAME: mma.sync.aligned.m16n8k8.row.col.f32.tf32.tf32.f32
// CHECK: llvm.inline_asm
// CHECK-SAME: mma.sync.aligned.m16n8k8.row.col.f32.tf32.tf32.f32
%28 = tt.dot %a_mat, %b_mat, %cst {allowTF32 = true, transA = false, transB = false} : tensor<32x16xf32, #dot_operand_a> * tensor<16x32xf32, #dot_operand_b> -> tensor<32x32xf32, #mma>
%38 = triton_gpu.convert_layout %28 : (tensor<32x32xf32, #mma>) -> tensor<32x32xf32, #blocked>
%30 = tt.splat %ptr : (!tt.ptr<f32>) -> tensor<32x1x!tt.ptr<f32>, #blocked>
%36 = tt.broadcast %30 : (tensor<32x1x!tt.ptr<f32>, #blocked>) -> tensor<32x32x!tt.ptr<f32>, #blocked>
tt.store %36, %38 : tensor<32x32xf32, #blocked>
return
}
}
// -----
#blocked0 = #triton_gpu.blocked<{sizePerThread = [1], threadsPerWarp = [32], warpsPerCTA = [4], order = [0]}>
module attributes {"triton_gpu.num-warps" = 4 : i32} {
// CHECK-LABEL: atomic_add_f32